Low Complexity Convolutional Neural Networks for Wireless Receiver Chain Optimization

Abstract

Technologies for 5G and beyond open up new chances for enabling new applications, which leads to an increasing variety of requirements, possible scenarios, and possible engineering decisions for wireless systems. Thus, having dynamic and robust techniques that can adapt to this huge variety has become more important than ever. One of the challenging adaptations is to select the most appropriate receiver architecture i.e. the architecture that gives the required performance with the least possible complexity, while modifying it dynamically based on the effects of having an instantaneous mix of a data sequence, channel effects, noise, and transmitter/receiver chains imperfections and impairments. One of the most innovative techniques is using convolutional neural networks (CNNs) as an initial pre-process that is capable of predicting the best receiver architecture. The technique depends on using offline pre-trained CNNs that can classify every incoming packet dynamically and assign it to the most appropriate receiver architecture. The technique shows high performance and accuracy that leads to higher certainty of the required resources and processing time, and consequently, better scheduling for the processes in the available receiver architectures and processing elements. Despite that, the technique adds an extra complexity due to the added CNNs. Although CNNs operations as multiplications have lower complexity than the operations in the receiver blocks, the added complexity due to using the CNNs is so high that they lead to total higher complexity than just using a higher complexity receiver in many cases.Here we propose a low complexity approach that gives an equivalent performance of the state of the art technique. Our approach here depends on reducing the size of the used CNNs by introducing parts of the incoming packet to the input layers of the CNNs instead of introducing the whole packet as in literature state of art, which reduces the added complexity due to the CNNs while keeping the advantage of the pre-knowledge of the required resources and the corresponding processing time. We show different approaches of how to extract enough information from the packet without the need to use all of it as input for the CNN, and analyzing the performance for every approach; then showing the total complexity reduction due to our new proposal.